The open‑source LLaMA inference library now includes cooperative‑k DP4A GEMM support for q4_0, q4_k and q6_k formats.
NoteBy OossaPublished by Oossa: 1 min read
The ggml‑org/llama.cpp project released version b11741 on 2026‑10‑19. It adds new OpenCL kernels that use DP4A instructions to speed up matrix multiplications for three quantized formats (q4_0, q4_k, q6_k). The update also fixes an alignment issue and bundles pre‑built binaries for macOS (Apple Silicon and Intel) and several Ubuntu targets, including Vulkan builds.
Developers can run LLaMA models faster on OpenCL‑compatible GPUs, making local inference more practical.
Models
The Chinese firm MiniMax launched M3.1 Flash, a cheaper but stronger reasoning model, with open weights slated for Oct 23, 2026.
Models
Mistral AI unveiled Timestral, a family of time‑series models from 5 M to 209 M parameters that beats Google’s TimesFM‑3 on several benchmarks and runs on a CPU.
Models
Four Gemini 4 Argon AI agents explored a random Minecraft world, coordinated themselves, and killed the Ender Dragon without any human direction.
Oossa · Newsletter
The week in AI, explained
Every Monday: the stories worth knowing, in plain language. Free, no spam.
